Source　localization　for　near-field　narrowband　signal　is　an　important　topic　in　array　signal　processing.　Deep　neural　network　(DNN)　based　methods　are　data-driven　and　free　of　pre-assumptions　about　data　model　and　are　expected　to　learn　the　intricate　nonlinear　structure　in　large　data　sets.　This　paper　proposes　a　framework　of　DNN　where　a　regression　layer　is　utilized　to　address　the　problem　of　near-field　source　localization.　Unlike　previous　studies　in　which　DOA　estimation　is　modeled　as　a　classification　problem　and　have　a　relatively　low　resolution,　we　exploit　a　regression　model　and　aim　to　improve　the　estimation　accuracy.　In　the　training　stage,　we　propose　a　novel　form　of　feature　representation　to　take　full　advantage　of　the　convolution　networks.　In　addition,　the　architecture　of　deep　neural　networks　is　well　designed　taking　in　to　consideration　the　trade-off　between　the　expression　ability　and　under-training　risks.　The　simulation　results　show　that　the　proposed　approach　has　a　rather　high　validation　accuracy　with　a　high　resolution,　and　also　outperforms　some　conventional　methods　in　adverse　environments　such　as　low　signal　to　noise　ratio　(SNR)　or　small　number　of　snapshots.　©　2019　IEEE